Disease and Medicaid Expansion
EDA Presentation | Cause: Level 4 | Measure: Death | Metric: Number | Age: All buckets | Sex: M & F
Housekeeping
Load data
Level Shift
If there is a level shift in the number of deaths around expansion years, it might suggest the presence of a treatment effect.
Let’s compute the deaths for the year after expansion minus the deaths in the year before expansion.
How to read this?
First, we reindexed deaths relative to expansion year. E.g. for California, 2015, <5, Female: (Death in 2015/Death 2014)*100
Second, we compute the change in that index 1yr post minus 1yr prior expansion year.
Third, we average that change across all locations, ages, sexes for each cause.
Fourth, I plot that average along with 95% CI.
Let’s say the value on the chart is 2. That could Chronic Kidney disease. That could mean that on average 1yr before expansion deaths were at 99 (i.e. ~1% lower than expansion year), and on average 1yr after expansion deaths were at 101 (i.e. ~1% higher than expansion year). So, deaths rose 2ppts (relative to expansion year deaths) around expansion year.
Interpretation: there definitely a few causes of deaths where there was a difference in the deaths before vs after. You can hover over the graph above to see the names of the causes and amounts.
e.g. Biggest drop: Neonatal preterm birth Biggest gain: opioid use disorders
Are these statistically significant? Recall that the variation varies dramatically across causes of death.
Let’s plot the t stats
Interpretation: Several diseases do indeed have statistically significant changes around the expansion year.
But don’t get too excited. There are clear trends in these series.
I wonder if this is the type of EDA the students’ were missing. If their DiD approach didn’t explicitly detrend, they would have been tricked.
Trend Adjusted
Let’s adjust for the trend within each cause of death and within each age/sex/location unit.
This is tricky. We want as much data as possible to build the trend, but
- We don’t want to use the post expansion periods, because that would blunt the impact we’re trying to see
- We don’t want to go too far back in time because several of the diseases I’ve been spot checking appear to have time varying trends.
- We could pick calendar years (let’s say 2003-2013). That would be just before the first expansion year. We could use this for all locations regardless of when they were treated. That might avoid the early flat spots we see in the data and would keep all of the units on the same # of observations. The downside is that the forecasts will be less powerful for the states and were treated later in our sample. The other option is to use the same window size of 10yrs pre to each unit’s treatment date. Neither are great, but let’s try the latter.
Load the detrended data
The dots in blue suggest that the change around expansion year is statistically significant after accounting for trend.
CAVEATS CAVEATS CAVEATS….10yrs is small. Trends may have varied in the training period, etc…
The top candidates
I screened the database for the causes where the |t-stat| >5 (i.e. possibly chance of treatment effect).
I then produced plots of the reindex deaths by cause for Age = All Ages, Sex= Both across all locations, grouped by expansion year cohort.
I didn’t separate by age bucket/sex combo because I want to find an AVERAGE Treatment effect. I won’t be able to say much rigorous one just one unit of observation.
You can view the plots in individual HTML files. These are interactive, so you can click and hover.
Like before, you are looking for shifts in level or trend.